Non-record: MLX-Optimized 12L 416d with SmearGate + BigramHash (val_bpb=1.9011, Mac)#342
Open
adhyaay-karnwal wants to merge 2 commits intoopenai:mainfrom
Open
Non-record: MLX-Optimized 12L 416d with SmearGate + BigramHash (val_bpb=1.9011, Mac)#342adhyaay-karnwal wants to merge 2 commits intoopenai:mainfrom
adhyaay-karnwal wants to merge 2 commits intoopenai:mainfrom
Conversation
…Hash Non-record submission for OpenAI Parameter Golf challenge. Trained on MacBook Apple Silicon M4 Pro using MLX framework. Key techniques: - 12 layers (6 encoder + 6 decoder) - 416 model dimension, MLP 3x expansion - SmearGate for local context - BigramHash with 4096 buckets - FP16 embeddings with Muon optimizer + weight decay - U-Net skip connections Result: val_bpb = 1.9011 (500 iterations, undertrained) This submission demonstrates effective MLX optimization techniques and serves as a foundation for further H100 training.
adhyaay-karnwal
changed the title
…techniques - train_sota.py: New script with BigramHash(10240), WD=0.04, SWA - train_optimized.py: Updated with faster validation - train_breakthrough.py, train_breakthrough_v3.py: Experimental versions - New submission folder with README and submission.json Key improvements from research: - BigramHash(10240): 2.5x larger than previous 4096 - SWA with start_frac=0.4: Optimal per openai#1 submission - Muon WD=0.04: Higher than previous 0.02 - SmearGate: Proven technique from top submissions - MLP 3x expansion: relu^2 activation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Key Techniques
Architecture
Training Details
Notes
This is an undertrained model on MacBook. The same architecture with 3000+ iterations on H100s should achieve significantly better BPB (potentially 1.5-1.6 BPB based on findings and research completed). This submission demonstrates effective MLX optimization techniques and serves as a foundation for further H100 training.
Files
records/track_non_record_16mb/2026-03-21_MLX_Optimized_12L_416d_SmearGate_BigramHash/README.md- Detailed explanationrecords/track_non_record_16mb/2026-03-21_MLX_Optimized_12L_416d_SmearGate_BigramHash/submission.json- Metadatarecords/track_non_record_16mb/2026-03-21_MLX_Optimized_12L_416d_SmearGate_BigramHash/train_gpt_mlx.py- MLX training scriptrecords/track_non_record_16mb/2026-03-21_MLX_Optimized_12L_416d_SmearGate_BigramHash/train.log- Training log